LORE_sa

This is the official repository of the LORE (Local Rule-Based Explanation) algorithm.

Getting started

We suggest to install the library and its requirements into a dedicated environment.

virtualenv venv
source venv/bin/activate
pip install -r requirements.txt

To use the library within your project just import the needed packages:

from lore_sa.dataset import TabularDataset

# load the training data
dataset = TabularDataset.from_csv('my_data.csv', class_name = "class")

...

Your first LORE explanation

Let's consider a simple example to explain the prediction of a model on a tabular dataset. We have a tabular dataset containing observation of a Credit Risk use case. We will create an opaque model using a Random Forest classifier and then we will use LORE to explain the prediction of the model on a specific instance.

Let's start by loading the dataset and creating the model:

import pandas as pd
from sklearn.ensemble import RandomForestClassifier
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import StandardScaler, OrdinalEncoder
from sklearn.pipeline import make_pipeline
from sklearn.model_selection import train_test_split

from lore_sa import sklearn_classifier_bbox

df = pd.read_csv('data/credit_risk.csv')
preprocessor = ColumnTransformer(
    transformers=[
        ('num', StandardScaler(), [0,8,9,10]),
        ('cat', OrdinalEncoder(), [1,2,3,4,5,6,7,11])
    ]
)
model = make_pipeline(preprocessor, RandomForestClassifier(n_estimators=100, random_state=42))

X_train, X_test, y_train, y_test = train_test_split(df.loc[:, 'age':'native-country'].values, df['class'].values,
                                            test_size=0.3, random_state=42, stratify=df['class'].values)
model.fit(X_train, y_train)

bbox = sklearn_classifier_bbox.sklearnBBox(model)

We wrap the model into a sklearnBBox object, which is a wrapper that allows LORE to interact with the model. Now we can use LORE to explain the prediction of the model on a specific instance. We need to provide LORE with information on the internal structure of the dataset. We will use the TabularDataset class to do so.

from lore_sa.dataset import TabularDataset
from lore_sa.generator import GeneticGenerator
from lore_sa.encoder_decoder import ColumnTransformerEnc
from lore_sa.lore import Lore
from lore_sa.surrogate import DecisionTreeSurrogate

dataset = TabularDataset.from_csv('resources/adult.csv', class_name='class')
dataset.df.dropna(inplace=True)
dataset.df.drop(['fnlwgt', 'education-num'], axis=1, inplace=True)
dataset.update_descriptor()

enc = ColumnTransformerEnc(dataset.descriptor)
generator = GeneticGenerator(bbox, dataset, enc)
surrogate = DecisionTreeSurrogate()

tabularLore = Lore(bbox, dataset, encoder, generator, surrogate)

Now we have an instance of the Lore class that we can use to explain the prediction of the model on a specific instance. Let's consider the first instance of the test set and explain the prediction of the model on this instance.

instance = X_test[0]
explanation = tabularLore.explain_instance(instance)
print(explanation)

Issue tracking

For any issue or bug, please open a new issue in the issue tracker available at: https://github.com/kdd-lab/LORE_sa/issues

Contributing

If you want to contribute to the library, please fork the repository and submit a pull request with the changes. The pull request will be reviewed by the maintainers and merged into the main branch if the changes are considered appropriate.

Documentation

The documentation is based on Sphinx. Documentation of the code is created by simply writing docstrings using reStructuredText markup. Docstrings are comments placed within triple quotes (''' or """) immediately below module, class, function, or method definitions.

The creation of online documentation the features of Sphinx. To build the documentation:

cd docs
make html

Once the documentation is built, the new folder docs/html must be committed and pushed to the repository and the documentation is then available here: https://kdd-lab.github.io/LORE_sa/html/index.html

To update the online documentation, as an instance when new modules or function are added to the LORE_sa library, it is necessary to delete the old folder docs/html, build the documentation (see the snippet above) and copy the greshly created docs/_build/html folder into docs/. Then, after committing and pushing the folder docs/html, the online documentation is updated to the last version.

kdd-lab / LORE_sa